Redis

The Redis facts that come up in backend and system design interviews, for Redis 7.4 and 8.x. Valkey, the BSD-licensed Linux Foundation fork of Redis 7.2.4, matches everything here except the modules built into Redis 8.

Core model

• In-memory server of typed values under string keys; commands run server-side, so each command is atomic.
• One main thread executes commands in an event loop (epoll/kqueue). Background threads free memory (UNLINK) and fsync the AOF; fork() makes RDB snapshots and AOF rewrites; io-threads N offloads socket reads, parsing and replies.
• Consequence: single commands never race, but one slow command (KEYS *, DEL of a huge key, a long Lua script) blocks every client.
• Licence: RSALv2/SSPLv1 since 7.4, AGPLv3 added as an option in Redis 8. Redis 8 bundles JSON, the Query Engine, time series, probabilistic types (Bloom, Cuckoo, Top-K) and vector sets.

Data types

• LPUSH l a b c then LRANGE l 0 -1 gives c b a. SADD returns how many members were new.
• Sorted-set ties sort by member name. ZREVRANK is zero-based from the top.

Expiry and eviction

SET k v EX 60          # TTL in seconds (PX for ms)
SET k v2 KEEPTTL       # plain SET would remove the TTL
EXPIRE k 120 GT        # NX | XX | GT | LT options (7.0+)
TTL k                  # -1 = no expiry, -2 = no such key
PERSIST k

• Expired keys are removed lazily (on access) and actively (a sampling cycle, hz 10).
• INCR, HSET, LPUSH, RENAME keep the TTL; SET, GETSET and DEL clear it. Replicas wait for the primary’s DEL.

• volatile-* with no TTL keys behaves like noeviction.
• LRU and LFU are approximate: maxmemory-samples 5 plus a 16-entry candidate pool. LFU uses a logarithmic 8-bit counter (lfu-log-factor 10) that decays (lfu-decay-time 1). OBJECT FREQ needs an LFU policy.
• Watch evicted_keys, expired_keys, keyspace_hits/keyspace_misses in INFO stats.

Caching patterns

• Invalidate with update DB, then delete; never update the cached value from the write path (writer races). TTL is the safety net; CDC or an outbox makes invalidation reliable.
• Stampede (hot key expires, everyone rebuilds): rebuild lock with SET NX PX, stale-while-revalidate, probabilistic early refresh, TTL jitter.
• Penetration (missing IDs): cache a short negative entry, Bloom filter. Avalanche (mass expiry or outage): jitter, HA, DB protection.

Persistence

• fork() costs page-table copy time (latest_fork_usec) and copy-on-write memory; disable transparent huge pages, set vm.overcommit_memory=1, keep headroom.
• Never auto-restart an empty primary with persistence off: replicas sync to the empty dataset.

Replication, Sentinel, Cluster

• REPLICAOF host port; PSYNC with replication ID + offset; partial resync from repl-backlog-size (1 MB default) or full resync (diskless by default since 7.0).
• Replication is asynchronous: acknowledged writes can vanish on failover. WAIT n ms, WAITAOF (7.2) and min-replicas-to-write narrow the window only.
• Sentinel: ≥ 3 processes; SDOWN after down-after-milliseconds, ODOWN when quorum agree, failover needs a majority leader vote. Clients ask SENTINEL get-master-addr-by-name.
• Cluster: 16,384 slots, slot = CRC16(key) mod 16384 (foo → 12182); bus port = data port + 10000; only DB 0.

• Hash tag: only the text inside the first {…} is hashed; {user:1000}:cart and {user:1000}:profile share a slot. Never put everything under one tag.

Atomicity tools

• Queue-time error (unknown command) → EXECABORT, nothing runs. Runtime error (WRONGTYPE) → that command fails, the others apply.
• Lua: pass keys in KEYS, values in ARGV; Lua number → integer (3.99 → 3); false → nil. Scripts over busy-reply-threshold (5 s) make others get BUSY.

Locks and rate limits

-- acquire: SET lock:job <token> NX PX 30000   (reply OK or nil)
-- release only if we still own it:
if redis.call('GET', KEYS[1]) == ARGV[1] then
  return redis.call('DEL', KEYS[1])
end
return 0

• Not SETNX + EXPIRE (crash in between leaves a lock forever); not a bare DEL (removes someone else’s lock after expiry).
• Pauses can outlive the TTL: use fencing tokens at the resource for correctness. Redlock = majority of 5 independent primaries; Kleppmann’s critique targets timing assumptions.

Messaging

• Pub/Sub: no storage, at-most-once; offline subscribers miss messages; slow ones hit client-output-buffer-limit pubsub 32mb 8mb 60. SPUBLISH (7.0) for sharded channels.
• Streams: IDs ms-seq, consumer groups, pending entries list until XACK, XAUTOCLAIM for dead consumers, trim with MAXLEN ~.

Operations and memory

• Compact encodings: hashes and sorted sets stay listpack up to 128 entries / 64-byte values; sets stay intset up to 512 integers. Strings: int, embstr (≤ 44 bytes), raw. Check with OBJECT ENCODING.

Practise with the full question bank on the Redis topic page.